CPU-Only & High-Memory Workloads

Not all scientific computations require GPUs. For high-memory genome assembly, chemistry simulations, or heavy data wrangling, Berkelium features high-core-count and large-memory worker nodes:

• Large-Memory Compute Nodes: Up to 224 CPU cores and 1–2 TB DDR5 ECC RAM per node.
• Standard High-Memory Nodes: 64+ CPU cores with 500+ GiB RAM and multi-terabyte local NVMe scratch per node.
• General CPU Nodes: 64 CPU cores for general services, preprocessing, and data ingest.

Check the live Cluster Status page for current node inventory and real-time resource availability.

---

Example: High-Memory Simulation Manifest

Specify CPU & Memory to Avoid Default LimitRange Throttling

If you omit resources.requests and resources.limits, the namespace default LimitRange will automatically assign fallback values (1 CPU limit, 100m CPU request, and 1Gi Memory). For CPU-intensive scientific simulations, this will severely throttle multi-threaded execution and cause immediate Out-Of-Memory (OOMKilled) termination if RAM exceeds 1Gi. Always define explicit resources!

Strictly Prohibited: Jobs Must Exit Automatically (No sleep infinity)

Batch Jobs must run to completion and exit by themselves without human intervention. Running sleep infinity, tail -f /dev/null, or equivalent commands to keep compute nodes reserved is strictly prohibited.

apiVersion: batch/v1
kind: Job
metadata:
  name: highmem-genome-assembly
  namespace: my-project
spec:
  template:
    spec:
      restartPolicy: OnFailure
      containers:
        - name: assembler
          image: biocontainers/spades:v3.15.5_cv1
          command: ["spades.py", "-1", "/data/reads_R1.fq", "-2", "/data/reads_R2.fq", "-o", "/data/assembly", "-t", "32", "-m", "250"]
          volumeMounts:
            - name: dataset
              mountPath: /data
          resources:
            requests:
              cpu: "16"
              memory: "128Gi"
            limits:
              cpu: "32"
              memory: "256Gi"
      volumes:
        - name: dataset
          persistentVolumeClaim:
            claimName: lab-shared-data-pvc

---

OpenMP & Thread Pinning

When running multi-threaded scientific binaries, ensure OpenMP threads match your requested CPU limits:

export OMP_NUM_THREADS=${OMP_NUM_THREADS:-$(nproc)}
export MKL_NUM_THREADS=$OMP_NUM_THREADS